Detecting Biomedical Copy-Move Forgery by Attention-Based Multiscale Deep Descriptors

Abstract

Translate Article

Continual revelations of academic fraud have raised concerns regarding the detection of forged experimental images in the public domain. To address this issue, we introduce a multiscale attention-based deep (MAD) descriptor scheme for detecting copy-move image forgery in biomedical research scenarios. Our method utilizes the common object detection network as a backbone and incorporates the positional embedding module, the channel-attention module, and the self-attention module to generate a dense feature field for input images. The proposed method demonstrates robustness against common attacks encountered during research manuscript preparation, low-contrast biomedical images featuring small foreground objects, and unseen or unlearned object patterns. Extensive experiments substantiate that our approach outperforms previous copy-move forgery methods when applied to real-world cases across various domains. Our proposed method can serve as an efficient screening tool for the rapid identification of biomedical image forgeries.